IB/ipath: Fix lost UD send work request
[safe/jmp/linux-2.6] / drivers / infiniband / hw / ipath / ipath_ud.c
1 /*
2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <rdma/ib_smi.h>
35
36 #include "ipath_verbs.h"
37 #include "ipath_kernel.h"
38
39 /**
40  * ipath_ud_loopback - handle send on loopback QPs
41  * @sqp: the sending QP
42  * @swqe: the send work request
43  *
44  * This is called from ipath_make_ud_req() to forward a WQE addressed
45  * to the same HCA.
46  * Note that the receive interrupt handler may be calling ipath_ud_rcv()
47  * while this is being called.
48  */
49 static void ipath_ud_loopback(struct ipath_qp *sqp, struct ipath_swqe *swqe)
50 {
51         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
52         struct ipath_qp *qp;
53         struct ib_ah_attr *ah_attr;
54         unsigned long flags;
55         struct ipath_rq *rq;
56         struct ipath_srq *srq;
57         struct ipath_sge_state rsge;
58         struct ipath_sge *sge;
59         struct ipath_rwq *wq;
60         struct ipath_rwqe *wqe;
61         void (*handler)(struct ib_event *, void *);
62         struct ib_wc wc;
63         u32 tail;
64         u32 rlen;
65         u32 length;
66
67         qp = ipath_lookup_qpn(&dev->qp_table, swqe->wr.wr.ud.remote_qpn);
68         if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
69                 dev->n_pkt_drops++;
70                 goto done;
71         }
72
73         rsge.sg_list = NULL;
74
75         /*
76          * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
77          * Qkeys with the high order bit set mean use the
78          * qkey from the QP context instead of the WR (see 10.2.5).
79          */
80         if (unlikely(qp->ibqp.qp_num &&
81                      ((int) swqe->wr.wr.ud.remote_qkey < 0 ?
82                       sqp->qkey : swqe->wr.wr.ud.remote_qkey) != qp->qkey)) {
83                 /* XXX OK to lose a count once in a while. */
84                 dev->qkey_violations++;
85                 dev->n_pkt_drops++;
86                 goto drop;
87         }
88
89         /*
90          * A GRH is expected to preceed the data even if not
91          * present on the wire.
92          */
93         length = swqe->length;
94         memset(&wc, 0, sizeof wc);
95         wc.byte_len = length + sizeof(struct ib_grh);
96
97         if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
98                 wc.wc_flags = IB_WC_WITH_IMM;
99                 wc.ex.imm_data = swqe->wr.ex.imm_data;
100         }
101
102         /*
103          * This would be a lot simpler if we could call ipath_get_rwqe()
104          * but that uses state that the receive interrupt handler uses
105          * so we would need to lock out receive interrupts while doing
106          * local loopback.
107          */
108         if (qp->ibqp.srq) {
109                 srq = to_isrq(qp->ibqp.srq);
110                 handler = srq->ibsrq.event_handler;
111                 rq = &srq->rq;
112         } else {
113                 srq = NULL;
114                 handler = NULL;
115                 rq = &qp->r_rq;
116         }
117
118         if (rq->max_sge > 1) {
119                 /*
120                  * XXX We could use GFP_KERNEL if ipath_do_send()
121                  * was always called from the tasklet instead of
122                  * from ipath_post_send().
123                  */
124                 rsge.sg_list = kmalloc((rq->max_sge - 1) *
125                                         sizeof(struct ipath_sge),
126                                        GFP_ATOMIC);
127                 if (!rsge.sg_list) {
128                         dev->n_pkt_drops++;
129                         goto drop;
130                 }
131         }
132
133         /*
134          * Get the next work request entry to find where to put the data.
135          * Note that it is safe to drop the lock after changing rq->tail
136          * since ipath_post_receive() won't fill the empty slot.
137          */
138         spin_lock_irqsave(&rq->lock, flags);
139         wq = rq->wq;
140         tail = wq->tail;
141         /* Validate tail before using it since it is user writable. */
142         if (tail >= rq->size)
143                 tail = 0;
144         if (unlikely(tail == wq->head)) {
145                 spin_unlock_irqrestore(&rq->lock, flags);
146                 dev->n_pkt_drops++;
147                 goto drop;
148         }
149         wqe = get_rwqe_ptr(rq, tail);
150         if (!ipath_init_sge(qp, wqe, &rlen, &rsge)) {
151                 spin_unlock_irqrestore(&rq->lock, flags);
152                 dev->n_pkt_drops++;
153                 goto drop;
154         }
155         /* Silently drop packets which are too big. */
156         if (wc.byte_len > rlen) {
157                 spin_unlock_irqrestore(&rq->lock, flags);
158                 dev->n_pkt_drops++;
159                 goto drop;
160         }
161         if (++tail >= rq->size)
162                 tail = 0;
163         wq->tail = tail;
164         wc.wr_id = wqe->wr_id;
165         if (handler) {
166                 u32 n;
167
168                 /*
169                  * validate head pointer value and compute
170                  * the number of remaining WQEs.
171                  */
172                 n = wq->head;
173                 if (n >= rq->size)
174                         n = 0;
175                 if (n < tail)
176                         n += rq->size - tail;
177                 else
178                         n -= tail;
179                 if (n < srq->limit) {
180                         struct ib_event ev;
181
182                         srq->limit = 0;
183                         spin_unlock_irqrestore(&rq->lock, flags);
184                         ev.device = qp->ibqp.device;
185                         ev.element.srq = qp->ibqp.srq;
186                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
187                         handler(&ev, srq->ibsrq.srq_context);
188                 } else
189                         spin_unlock_irqrestore(&rq->lock, flags);
190         } else
191                 spin_unlock_irqrestore(&rq->lock, flags);
192
193         ah_attr = &to_iah(swqe->wr.wr.ud.ah)->attr;
194         if (ah_attr->ah_flags & IB_AH_GRH) {
195                 ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
196                 wc.wc_flags |= IB_WC_GRH;
197         } else
198                 ipath_skip_sge(&rsge, sizeof(struct ib_grh));
199         sge = swqe->sg_list;
200         while (length) {
201                 u32 len = sge->length;
202
203                 if (len > length)
204                         len = length;
205                 if (len > sge->sge_length)
206                         len = sge->sge_length;
207                 BUG_ON(len == 0);
208                 ipath_copy_sge(&rsge, sge->vaddr, len);
209                 sge->vaddr += len;
210                 sge->length -= len;
211                 sge->sge_length -= len;
212                 if (sge->sge_length == 0) {
213                         if (--swqe->wr.num_sge)
214                                 sge++;
215                 } else if (sge->length == 0 && sge->mr != NULL) {
216                         if (++sge->n >= IPATH_SEGSZ) {
217                                 if (++sge->m >= sge->mr->mapsz)
218                                         break;
219                                 sge->n = 0;
220                         }
221                         sge->vaddr =
222                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
223                         sge->length =
224                                 sge->mr->map[sge->m]->segs[sge->n].length;
225                 }
226                 length -= len;
227         }
228         wc.status = IB_WC_SUCCESS;
229         wc.opcode = IB_WC_RECV;
230         wc.qp = &qp->ibqp;
231         wc.src_qp = sqp->ibqp.qp_num;
232         /* XXX do we know which pkey matched? Only needed for GSI. */
233         wc.pkey_index = 0;
234         wc.slid = dev->dd->ipath_lid |
235                 (ah_attr->src_path_bits &
236                  ((1 << dev->dd->ipath_lmc) - 1));
237         wc.sl = ah_attr->sl;
238         wc.dlid_path_bits =
239                 ah_attr->dlid & ((1 << dev->dd->ipath_lmc) - 1);
240         wc.port_num = 1;
241         /* Signal completion event if the solicited bit is set. */
242         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
243                        swqe->wr.send_flags & IB_SEND_SOLICITED);
244 drop:
245         kfree(rsge.sg_list);
246         if (atomic_dec_and_test(&qp->refcount))
247                 wake_up(&qp->wait);
248 done:;
249 }
250
251 /**
252  * ipath_make_ud_req - construct a UD request packet
253  * @qp: the QP
254  *
255  * Return 1 if constructed; otherwise, return 0.
256  */
257 int ipath_make_ud_req(struct ipath_qp *qp)
258 {
259         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
260         struct ipath_other_headers *ohdr;
261         struct ib_ah_attr *ah_attr;
262         struct ipath_swqe *wqe;
263         unsigned long flags;
264         u32 nwords;
265         u32 extra_bytes;
266         u32 bth0;
267         u16 lrh0;
268         u16 lid;
269         int ret = 0;
270         int next_cur;
271
272         spin_lock_irqsave(&qp->s_lock, flags);
273
274         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
275                 if (!(ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND))
276                         goto bail;
277                 /* We are in the error state, flush the work request. */
278                 if (qp->s_last == qp->s_head)
279                         goto bail;
280                 /* If DMAs are in progress, we can't flush immediately. */
281                 if (atomic_read(&qp->s_dma_busy)) {
282                         qp->s_flags |= IPATH_S_WAIT_DMA;
283                         goto bail;
284                 }
285                 wqe = get_swqe_ptr(qp, qp->s_last);
286                 ipath_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
287                 goto done;
288         }
289
290         if (qp->s_cur == qp->s_head)
291                 goto bail;
292
293         wqe = get_swqe_ptr(qp, qp->s_cur);
294         next_cur = qp->s_cur + 1;
295         if (next_cur >= qp->s_size)
296                 next_cur = 0;
297
298         /* Construct the header. */
299         ah_attr = &to_iah(wqe->wr.wr.ud.ah)->attr;
300         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
301                 if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
302                         dev->n_multicast_xmit++;
303                 else
304                         dev->n_unicast_xmit++;
305         } else {
306                 dev->n_unicast_xmit++;
307                 lid = ah_attr->dlid & ~((1 << dev->dd->ipath_lmc) - 1);
308                 if (unlikely(lid == dev->dd->ipath_lid)) {
309                         /*
310                          * If DMAs are in progress, we can't generate
311                          * a completion for the loopback packet since
312                          * it would be out of order.
313                          * XXX Instead of waiting, we could queue a
314                          * zero length descriptor so we get a callback.
315                          */
316                         if (atomic_read(&qp->s_dma_busy)) {
317                                 qp->s_flags |= IPATH_S_WAIT_DMA;
318                                 goto bail;
319                         }
320                         qp->s_cur = next_cur;
321                         spin_unlock_irqrestore(&qp->s_lock, flags);
322                         ipath_ud_loopback(qp, wqe);
323                         spin_lock_irqsave(&qp->s_lock, flags);
324                         ipath_send_complete(qp, wqe, IB_WC_SUCCESS);
325                         goto done;
326                 }
327         }
328
329         qp->s_cur = next_cur;
330         extra_bytes = -wqe->length & 3;
331         nwords = (wqe->length + extra_bytes) >> 2;
332
333         /* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
334         qp->s_hdrwords = 7;
335         qp->s_cur_size = wqe->length;
336         qp->s_cur_sge = &qp->s_sge;
337         qp->s_dmult = ah_attr->static_rate;
338         qp->s_wqe = wqe;
339         qp->s_sge.sge = wqe->sg_list[0];
340         qp->s_sge.sg_list = wqe->sg_list + 1;
341         qp->s_sge.num_sge = wqe->wr.num_sge;
342
343         if (ah_attr->ah_flags & IB_AH_GRH) {
344                 /* Header size in 32-bit words. */
345                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
346                                                  &ah_attr->grh,
347                                                  qp->s_hdrwords, nwords);
348                 lrh0 = IPATH_LRH_GRH;
349                 ohdr = &qp->s_hdr.u.l.oth;
350                 /*
351                  * Don't worry about sending to locally attached multicast
352                  * QPs.  It is unspecified by the spec. what happens.
353                  */
354         } else {
355                 /* Header size in 32-bit words. */
356                 lrh0 = IPATH_LRH_BTH;
357                 ohdr = &qp->s_hdr.u.oth;
358         }
359         if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
360                 qp->s_hdrwords++;
361                 ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
362                 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
363         } else
364                 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
365         lrh0 |= ah_attr->sl << 4;
366         if (qp->ibqp.qp_type == IB_QPT_SMI)
367                 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
368         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
369         qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid);  /* DEST LID */
370         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
371                                            SIZE_OF_CRC);
372         lid = dev->dd->ipath_lid;
373         if (lid) {
374                 lid |= ah_attr->src_path_bits &
375                         ((1 << dev->dd->ipath_lmc) - 1);
376                 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
377         } else
378                 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
379         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
380                 bth0 |= 1 << 23;
381         bth0 |= extra_bytes << 20;
382         bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
383                 ipath_get_pkey(dev->dd, qp->s_pkey_index);
384         ohdr->bth[0] = cpu_to_be32(bth0);
385         /*
386          * Use the multicast QP if the destination LID is a multicast LID.
387          */
388         ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
389                 ah_attr->dlid != IPATH_PERMISSIVE_LID ?
390                 __constant_cpu_to_be32(IPATH_MULTICAST_QPN) :
391                 cpu_to_be32(wqe->wr.wr.ud.remote_qpn);
392         ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
393         /*
394          * Qkeys with the high order bit set mean use the
395          * qkey from the QP context instead of the WR (see 10.2.5).
396          */
397         ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->wr.wr.ud.remote_qkey < 0 ?
398                                          qp->qkey : wqe->wr.wr.ud.remote_qkey);
399         ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
400
401 done:
402         ret = 1;
403         goto unlock;
404
405 bail:
406         qp->s_flags &= ~IPATH_S_BUSY;
407 unlock:
408         spin_unlock_irqrestore(&qp->s_lock, flags);
409         return ret;
410 }
411
412 /**
413  * ipath_ud_rcv - receive an incoming UD packet
414  * @dev: the device the packet came in on
415  * @hdr: the packet header
416  * @has_grh: true if the packet has a GRH
417  * @data: the packet data
418  * @tlen: the packet length
419  * @qp: the QP the packet came on
420  *
421  * This is called from ipath_qp_rcv() to process an incoming UD packet
422  * for the given QP.
423  * Called at interrupt level.
424  */
425 void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
426                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
427 {
428         struct ipath_other_headers *ohdr;
429         int opcode;
430         u32 hdrsize;
431         u32 pad;
432         struct ib_wc wc;
433         u32 qkey;
434         u32 src_qp;
435         u16 dlid;
436         int header_in_data;
437
438         /* Check for GRH */
439         if (!has_grh) {
440                 ohdr = &hdr->u.oth;
441                 hdrsize = 8 + 12 + 8;   /* LRH + BTH + DETH */
442                 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
443                 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
444                 header_in_data = 0;
445         } else {
446                 ohdr = &hdr->u.l.oth;
447                 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
448                 /*
449                  * The header with GRH is 68 bytes and the core driver sets
450                  * the eager header buffer size to 56 bytes so the last 12
451                  * bytes of the IB header is in the data buffer.
452                  */
453                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
454                 if (header_in_data) {
455                         qkey = be32_to_cpu(((__be32 *) data)[1]);
456                         src_qp = be32_to_cpu(((__be32 *) data)[2]);
457                         data += 12;
458                 } else {
459                         qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
460                         src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
461                 }
462         }
463         src_qp &= IPATH_QPN_MASK;
464
465         /*
466          * Check that the permissive LID is only used on QP0
467          * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
468          */
469         if (qp->ibqp.qp_num) {
470                 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
471                              hdr->lrh[3] == IB_LID_PERMISSIVE)) {
472                         dev->n_pkt_drops++;
473                         goto bail;
474                 }
475                 if (unlikely(qkey != qp->qkey)) {
476                         /* XXX OK to lose a count once in a while. */
477                         dev->qkey_violations++;
478                         dev->n_pkt_drops++;
479                         goto bail;
480                 }
481         } else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
482                    hdr->lrh[3] == IB_LID_PERMISSIVE) {
483                 struct ib_smp *smp = (struct ib_smp *) data;
484
485                 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
486                         dev->n_pkt_drops++;
487                         goto bail;
488                 }
489         }
490
491         /*
492          * The opcode is in the low byte when its in network order
493          * (top byte when in host order).
494          */
495         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
496         if (qp->ibqp.qp_num > 1 &&
497             opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
498                 if (header_in_data) {
499                         wc.ex.imm_data = *(__be32 *) data;
500                         data += sizeof(__be32);
501                 } else
502                         wc.ex.imm_data = ohdr->u.ud.imm_data;
503                 wc.wc_flags = IB_WC_WITH_IMM;
504                 hdrsize += sizeof(u32);
505         } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
506                 wc.ex.imm_data = 0;
507                 wc.wc_flags = 0;
508         } else {
509                 dev->n_pkt_drops++;
510                 goto bail;
511         }
512
513         /* Get the number of bytes the message was padded by. */
514         pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
515         if (unlikely(tlen < (hdrsize + pad + 4))) {
516                 /* Drop incomplete packets. */
517                 dev->n_pkt_drops++;
518                 goto bail;
519         }
520         tlen -= hdrsize + pad + 4;
521
522         /* Drop invalid MAD packets (see 13.5.3.1). */
523         if (unlikely((qp->ibqp.qp_num == 0 &&
524                       (tlen != 256 ||
525                        (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
526                      (qp->ibqp.qp_num == 1 &&
527                       (tlen != 256 ||
528                        (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
529                 dev->n_pkt_drops++;
530                 goto bail;
531         }
532
533         /*
534          * A GRH is expected to preceed the data even if not
535          * present on the wire.
536          */
537         wc.byte_len = tlen + sizeof(struct ib_grh);
538
539         /*
540          * Get the next work request entry to find where to put the data.
541          */
542         if (qp->r_flags & IPATH_R_REUSE_SGE)
543                 qp->r_flags &= ~IPATH_R_REUSE_SGE;
544         else if (!ipath_get_rwqe(qp, 0)) {
545                 /*
546                  * Count VL15 packets dropped due to no receive buffer.
547                  * Otherwise, count them as buffer overruns since usually,
548                  * the HW will be able to receive packets even if there are
549                  * no QPs with posted receive buffers.
550                  */
551                 if (qp->ibqp.qp_num == 0)
552                         dev->n_vl15_dropped++;
553                 else
554                         dev->rcv_errors++;
555                 goto bail;
556         }
557         /* Silently drop packets which are too big. */
558         if (wc.byte_len > qp->r_len) {
559                 qp->r_flags |= IPATH_R_REUSE_SGE;
560                 dev->n_pkt_drops++;
561                 goto bail;
562         }
563         if (has_grh) {
564                 ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
565                                sizeof(struct ib_grh));
566                 wc.wc_flags |= IB_WC_GRH;
567         } else
568                 ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
569         ipath_copy_sge(&qp->r_sge, data,
570                        wc.byte_len - sizeof(struct ib_grh));
571         if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
572                 goto bail;
573         wc.wr_id = qp->r_wr_id;
574         wc.status = IB_WC_SUCCESS;
575         wc.opcode = IB_WC_RECV;
576         wc.vendor_err = 0;
577         wc.qp = &qp->ibqp;
578         wc.src_qp = src_qp;
579         /* XXX do we know which pkey matched? Only needed for GSI. */
580         wc.pkey_index = 0;
581         wc.slid = be16_to_cpu(hdr->lrh[3]);
582         wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
583         dlid = be16_to_cpu(hdr->lrh[1]);
584         /*
585          * Save the LMC lower bits if the destination LID is a unicast LID.
586          */
587         wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
588                 dlid & ((1 << dev->dd->ipath_lmc) - 1);
589         wc.port_num = 1;
590         /* Signal completion event if the solicited bit is set. */
591         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
592                        (ohdr->bth[0] &
593                         __constant_cpu_to_be32(1 << 23)) != 0);
594
595 bail:;
596 }